# Give a bounded task and check the result

Give RX one project outcome and define how you will judge it. This guide concerns task scope and review; the prompt is not an enforcement policy.

## Before you submit

Open the intended [workspace](/docs/agent-rx/quickstart), save unrelated work, and identify a relevant project check. Confirm the provider route and account access. Submitting a task may send context to the chosen provider, invoke tools, and incur that provider's charges.

## Write the brief

1. Name one outcome, the files or subsystem involved, and changes to preserve.
2. Name the acceptance check and actions requiring another decision. Use this document template as input to your task; do not execute it as a shell command.

   ```markdown
   # Task brief

   ## Goal

   Describe one outcome.

   ## Scope

   List the files or subsystem to inspect.

   ## Existing work

   State which changes must be preserved.

   ## Acceptance

   Name the relevant test or observable result.

   ## Approval boundary

   List actions that need a separate decision.
   ```

3. Fill every section with your project's facts. For example, limit a parser correction to the failing parser case, preserve unrelated UI edits, and request the project's existing parser test. Do not invent a test command if the project does not document one.
4. Submit the completed request and watch file and shell activity. If the proposed action exceeds your intended scope, stop and reassess before continuing.

## Review the result

Ask which files changed and which exact check ran. Then inspect the working tree yourself. Replace the sample path with the relevant file.

```bash
git status --short
git diff --stat
git diff -- path/to/changed-file
```

The [Git diff reference](https://git-scm.com/docs/git-diff) explains what comparisons show. Check untracked files separately; a clean tracked diff does not account for all files. Run or inspect the relevant check and retain its actual exit status and failure output. A proposed correction, changed file, passing test, and completed task are distinct results.

## If work is incomplete

Preserve the diff and non-secret failure. Narrow the next request to the unresolved case instead of replaying the original task blindly. Do not discard your own changes to obtain a clean tree. Read [tool activity](/docs/agent-rx/read-file-and-shell-tool-activity) when the response and the observed project state disagree.